icml2027_mg_archive: π0.5 RoboCasa365 pilot checkpoints (archived 2026-10-23)
These checkpoints were archived when the project paused. They are research checkpoints, not a release. The best atomic-only π0.5 model is in ducido/pi05_atomic65_50k (35000/, 67.6% mean SR on the 65 RoboCasa365 atomic tasks, 20 episodes per task).
Format: openpi params-only checkpoints (orbax params/) with assets/<asset_id>/norm_stats.json. They load with openpi CheckpointWeightLoader or serve_policy together with the matching config from the project's openpi fork.
| Folder | Config | What it is | Mean SR |
|---|---|---|---|
sd1_ctrl_r1_3999/ |
pi05_atomic65_sd1_ctrl |
Control arm. Fine-tune pi05_atomic65_50k/35000 for 4k steps (LR 5e-6 → 1e-6, batch 64) on all 65 human atomic datasets. The 12 pilot tasks get 2× exposure through duplicated human data. |
59.6% (12 pilot tasks × 20 ep, pretrain split); 62.5% (18 atomic_seen tasks × 20 ep, target split) |
sd1_roll_r1_3999/ |
pi05_atomic65_sd1_roll |
SimDeploy arm (T011). Same recipe; the extra exposure is 540 successful rollouts of the policy itself instead of duplicated human data. | 58.3% (12 × 20) |
sd1_aug_r1_3999/ |
pi05_atomic65_sd1_aug |
Visual-DR arm (T012, SimVLA-style). Same recipe; the extra exposure is human ×1 plus two domain-randomized re-renders of the same demos (texture, material, light, agent-view camera), 0.5× each. | 57.1% (12 × 20); 64.2% (18 × 20, target split) |
t002_b3_29999/ |
pi05_robocasa_finetune_target_composite_unseen_10p |
G0 baseline B3. RoboCasa365 Human300 π0.5 (75k) fine-tuned on 10% of the Composite-Unseen target demos (50 per task), step 29999. | 28.8% on Composite-Unseen |
The base checkpoint (35000) scores 58.3% on the same 12 pilot tasks × 20 ep. Neither pilot arm beat the control by the +5 point threshold. Eval used paired episode seeds and Mesa llvmpipe rendering with LP_NUM_THREADS=1.
Pilot tasks: CloseCabinet, CoffeeSetupMug, OpenDrawer, OpenFridge, PickPlaceCounterToCabinet, PickPlaceCounterToMicrowave, PickPlaceCounterToStove, PickPlaceSinkToCounter, PreheatOven, TurnOnMicrowave, TurnOnSinkFaucet, TurnOnStove.
Training data: ducido/icml2027_mg_archive_data has the DR re-renders and the SimDeploy rollouts.